Enhancing Federated Learning with Globally Shared Model: A Modified FedAVG Approach (GSM-FedAVG)
Oudarja Barman Tanmoy, Md. Al Mamun, Sakib Hasan, Adnan Anwar
Abstract
The wireless communications industry is interested in using data-driven machine learning solutions to supplement traditional model-driven design processes. Decentralized ML algorithms that maintain data in its original location are more alluring due to the lack of accessibility of private data and significant communication overhead. Federated learning is crucial for wireless applications to protect user privacy. The Federated Averaging Algorithm is a technique used in the Federated Learning process, which cooperatively learns an ML model by using the simple arithmetic average. In this paper, we propose a technique for enhancing the robustness of FedAVG, which involves employing a globally shared model with each clients in addition to each client’s local model. This method is referred to as the GSM-FedAVG approach. The GSM approach with FedAVG has shown promising results in improving the accuracy and robustness of the ML models used as global model in FL setting. The effectiveness of the proposed method is validated using a number of benchmark datasets.